What actually works in AI visibility, graded by evidence
Twenty interventions sold as AI visibility work, each rated by how strong the evidence behind it actually is. Three are gates you must pass. One has a rigorous causal study behind it, and that study found nothing.
Almost everything written about getting recommended by AI assistants is a list of tactics with no indication of which ones have evidence behind them. This guide is that list with the evidence attached, and with the grade stated before the claim rather than buried under it.
The grading is simple. Grade A means the AI provider documents it themselves, or a controlled reproducible study demonstrates it. Grade B means credible practitioner testing with the methodology shown. Grade C means somebody asserted it, usually somebody selling it.
One caveat matters more than any single row below. Nearly every headline figure in this field — the +31%, the +22%, the 3x — is correlational. It comes from observing which already-cited pages happen to have which features. It does not establish that adding the feature causes the citation. The single largest causal test anyone has run in this space returned a null result, and it is item 12.
The three gates
These are not growth levers. They are conditions. If one of them is shut, nothing else on the list can help you, and no amount of content will change that. They are also the only items on the list graded A on the positive side.
| # | Intervention | Grade | What the evidence says |
|---|---|---|---|
| 1 | Crawler access and Bing indexation | A | Provider documentation. ChatGPT's URL discovery runs through the Bing Search API, so Bing indexation is a hard gatekeeper. Google requires a page to be indexed and eligible to be shown with a snippet. Blocking OAI-SearchBot in robots.txt removes you from citation eligibility outright. |
| 2 | Server-side rendering, no JavaScript dependency | A/B | Vercel and MERJ observed 500 million GPTBot fetches and found zero evidence of JavaScript execution. If your content only renders client-side, retrieval cannot see it. Binary, verifiable, fixable. |
| 3 | Google Business Profile and listing consistency | A | Google explicitly names Google Business Profile and Merchant Center as inputs to its generative AI features. For a local business this is the highest-confidence lever available. |
Pass all three gates before paying anyone for anything else. They are cheap, they are checkable in an afternoon, and every one of them is documented by the provider rather than inferred by a vendor.
The growth levers, and why they are all softer than they look
These have the best evidence of anything that could plausibly move your visibility upward rather than merely stop it being zero. Every one of them is correlational, a proxy measurement, or both.
| # | Intervention | Grade | What the evidence says |
|---|---|---|---|
| 4 | Third-party mentions of your brand | B | Ahrefs measured 75,000 brands: unlinked web mentions correlate with AI visibility at r=0.664, against r=0.218 for total backlinks. Mentions correlate roughly three times more strongly than links. Correlational, and the direction of causation is genuinely ambiguous — large brands get mentioned and get cited. |
| 5 | YouTube presence | B | The strongest single correlation measured anywhere in this field, r=0.737, from the same 75,000-brand study. YouTube accounts for roughly 5.6% of citations. Plausibly a proxy for brand size rather than a lever. |
| 6 | Reddit and community presence | B | Reddit is the single most-cited domain across every major engine. One documented 14-week programme moved a brand from 1 of 12 to 4 of 12 shortlist prompts, with 2 of the 4 new appearances traceable to specific threads the team took part in. That is the closest thing to a causal trace in this literature. |
| 7 | Content freshness | B | 76.4% of the most-cited ChatGPT pages were updated within the previous 30 days. Correlational, but cheap to act on and essentially risk-free. |
| 8 | Answer-first structure, statistics, quotations | B | The GEO-bench study across 10,000 queries reported +22% to +41%. Downgraded from A because it ran against a simulated engine, the metric is a proxy rather than real-world recommendations, and 40% is a maximum rather than a mean. Widely misquoted as 'boosts AI citations 40%'. |
Item 6 deserves a note. Reddit being the most-cited domain is not permission to astroturf it. Posting promotional content under the guise of a personal recommendation breaks Reddit's rules and is a reputational problem that outlasts any visibility gain.
The one rigorous causal study, and its null result
Structured data — JSON-LD schema markup — is the most confidently sold intervention in this category. It is also the only one subjected to a properly designed causal test at scale.
Ahrefs ran a difference-in-differences study between August 2025 and March 2026: 1,885 pages that added schema, against roughly 4,000 matched control pages.
| Platform | Effect of adding schema |
|---|---|
| Google AI Overviews | −4.6% (a small but statistically significant decline) |
| Google AI Mode | +2.4% (indistinguishable from noise) |
| ChatGPT | +2.2% (indistinguishable from noise) |
“Adding schema produced no major uplift in citations on any platform.”
Google's own guidance agrees: structured data is not required for generative AI search. The corroborating mechanism is that AI systems appear to extract visible HTML during retrieval rather than reading JSON-LD.
Schema is worth having because it is correct, standard, and costs an hour. It is not worth paying a monthly retainer for, and anyone presenting it as a primary AI visibility lever is contradicted by the best-designed study in their own field.
One honest caveat on that study: its population was pages that were already heavily cited. It does not test what schema does for a page nobody has ever cited. That is a real limit, and it is why the recommendation is 'do it, don't buy it' rather than 'don't bother'.
What is sold on nothing
| # | Intervention | Grade | The problem |
|---|---|---|---|
| 14 | llms.txt | C | No major AI provider has committed to reading it in production retrieval. SE Ranking found no correlation with citations across 300,000 domains. Google states you do not need to create AI text files. Covered in full in its own guide. |
| 15 | Review volume as an AI lever | C | Plausible, because reviews feed Google Business Profile, which Google does name as an input. But no controlled study isolating review volume's effect on AI recommendation appears to exist. Sell it as local hygiene, not as AI evidence. |
| 16 | Vector embedding optimisation | C | Theoretically sound and practically unmeasurable. The actionable advice underneath the jargon is identical to ordinary topical authority. A premium charged for vocabulary. |
| 17 | Speakable schema | C | One unverified claim of +127% voice referrals, a single anecdote, no platform confirmation. |
| 20 | Getting into training data | C | Unfalsifiable by construction. Training composition is opaque; you cannot verify inclusion or measure influence. Any pitch resting on this cannot be checked, which is the point. |
Two things nobody should sell you
Serving different content to AI crawlers than to human visitors is cloaking. It is sold as 'edge delivery for crawlers' or 'AI shadow sites'. There is no public validation data and unclear penalty exposure. Prompt injection hidden in page content is worse: it is manipulation with legal and reputational exposure, and it exists only in academic papers.
The honest summary
Items 1 to 3 are gates. They are Grade A and verifiable, and they are technical hygiene rather than a product. Items 4 to 8 carry the best growth evidence and every one of them is correlational or proxy-based. Item 12, schema, is the most rigorous causal test in the field, and it found nothing.
That asymmetry is the single most useful thing to carry into a conversation with any agency, including this one: the better the study design in this field, the smaller the measured effect. Anyone selling certainty is selling ahead of the evidence.
Common questions
- What is the single most important thing for AI visibility?
- Passing the three gates: being in Bing's index, allowing OAI-SearchBot and the other AI crawlers in robots.txt, and rendering your content server-side. They are the only Grade A positive items on the list, and if one of them is shut nothing else can work.
- Does schema markup help you get cited by AI?
- The one large-scale causal test found no meaningful uplift on any platform, and a small decline on Google AI Overviews. Google's own guidance says structured data is not required for generative AI search. Add schema because it is correct and cheap, not because it will get you cited.
- Why are so many AI visibility statistics unreliable?
- Almost all of them are correlational. They come from observing which already-cited pages have which features, which cannot show that adding the feature caused the citation. Larger brands tend to have all the features and all the citations.
Sources
Every claim above should be checkable. Where a study has limits, they are stated rather than left out.
- Ahrefs — schema difference-in-differences study, 1,885 treated pages vs ~4,000 controlsGrade A. The only large-scale causal test located in this field. Null result.
- Ahrefs — 75,000 brands, Spearman correlations for mentions, backlinks and YouTubeGrade B. Correlational.
- Vercel and MERJ — 500 million GPTBot fetches, JavaScript execution analysisGrade A/B.
- GEO-bench (Princeton, Georgia Tech, IIT Delhi, Allen AI), KDD 2024, 10,000 queriesGrade B. Simulated engine, proxy metric.
- SE Ranking — llms.txt correlation across 300,000 domainsGrade B. No correlation found.
- Google Search Central — generative AI and structured data guidanceGrade A. Provider documentation.
- Perplexity Help Center — how Perplexity follows robots.txtGrade A. Provider documentation.
More on getting cited in ai answers
- Why the five AI assistants disagree about who to recommend
We measured every source five assistants used to answer the same questions. Not one domain was cited by all five, and 89% were cited by only one. Our own original data on why a single AI visibility number is misleading.
- The source ecosystem: how AI actually decides who to name
When we asked five assistants who to hire in our own category, 59 of 108 citations were other people's roundup lists. Five of the twelve most-named companies sat on one directory. Being named is mostly about other people's pages.
- How AI assistants pick which businesses to name
ChatGPT, Perplexity, Google's AI answers and Copilot use four different pipelines to decide who gets recommended. An intervention that works on one can do nothing on another, which is why they have to be measured separately.
- What to fix first for AI search
A running order based on what the evidence supports rather than what is easiest to sell. Three gates before anything else, then off-site work, then your own pages — which come later than almost every agency will tell you.
- Does llms.txt do anything?
Over 844,000 sites have adopted llms.txt. No major AI provider has committed to reading it in production retrieval, and the largest correlation study found no relationship with citations. Here is what it is actually for.
Want to know where you stand?
We ask five AI assistants for a company like yours and send you a free report showing how often you came up and who came up instead.
Get my free report